Efficient Generation of Review Summaries

1. A computing system comprising: a processor; and computer storage memory having computer-executable instructions stored thereon which, when executed by the processor, configure the computing system to perform operations comprising: obtain a plurality of reviews that each comprise a description associated with an item; determine a set of reviews, from the plurality of reviews, to include in a model prompt in accordance with an input prompt size constraint associated with a trained large language model, wherein the set of reviews are determined from the plurality of reviews based on dates associated with the plurality of reviews, applicability to a recent version of the item, or a particular attribute associated with reviewers of the set of reviews, such that the set of reviews when included the model prompt cause the trained large language model to generate an item insight based on content of the set of reviews; generate, based on the set of reviews, a data structure including information indicating the item, the description associated with the item, the particular attribute, and weights corresponding to reviews of the set of reviews; generate the model prompt to be input into the trained large language model, the model prompt including the data structure and the set of reviews; and obtain, as output from the trained large language model, a review summary that summarizes the set of reviews associated with the item and indicates the item insight.

2. The computing system of claim 1, wherein the plurality of reviews are authored based on prior experience with the item.

3. The computing system of claim 1, wherein the particular attribute further comprises weights associated with the plurality of reviews.

4. The computing system of claim 3, wherein the weights are based on obtained review feedback.

5. The computing system of claim 1 wherein the particular attribute further comprises a context associated with reviews of the plurality of reviews.

6. The computing system of claim 1, further comprising: storing the review summary that summarizes the set of reviews associated with the item; and in response to a request for the review summary, providing at least a portion of the review summary for display.

7. The computing system of claim 1, wherein the model prompt includes an output attribute indicating a desired output associated with the review summary.

8. The computing system of claim 7, wherein the output attribute comprises a target type, a length, or a target language.

9. A computer-implemented method comprising: obtaining a plurality of reviews associated with an item; filtering, based on an intent of a review summary, a set of reviews of the plurality of reviews, where the set of reviews are selected based on a context associated with the intent; generating a data structure including an indication of the item and the context based on the set of reviews generating a prompt to be inputted into a trained large language model, the prompt including the data structure and the set of reviews, such that the set of reviews when included in the prompt causes the trained large language model to generate an item insight based on content of the set of reviews; and obtaining, as output from the trained large language model, the review summary, including the intent, that summarizes the set of reviews and indicates an item insight.

10. The method of claim 9, wherein the context includes an item context associated with the item.

11. The method of claim 10, wherein the item context includes at least one of: a release date of the item, a version of the item, a publisher of the item, and metadata associated with the item.

12. The method of claim 9, wherein the context includes a review context associated with a first review of the set of reviews.

13. The method of claim 12, wherein the review context includes at least one of: a reviewer identifier, a date associated with the first review, a tone associated with the first review, demographics information associated with the first review, indication of purchases associated with the first review, and a feedback associated with the first review.

14. The method of claim 9, wherein the method further comprises assigning a weight to a first review of the set of reviews and including the weight to the prompt.

15. The method of claim 14, wherein the weight is generated by the trained large language model based on the first review.

16. The method of claim 9, wherein the intent further comprises at least one of: a positive item insight, a constructive item insight, and a negative item insight.

17. One or more computer storage media having computer-executable instructions embodied thereon that, when executed by one or more processors, cause the one or more processors to perform operations comprising: selecting, from a plurality of reviews including a description associated with an item, a review to include in a model prompt in accordance with an input size constraint associated with a trained model, wherein the review is selected based on at least one of a review context, an item context, user data, reviewer data, and a review weight; generating a data structure based on the plurality of reviews, the data structure including an indication of the item, the description associated with the item, an attribute associated with the review, and the review weight; generating a prompt to be input into the trained model, the prompt including the data structure and the review, such that the review when included in the prompt cause the trained model to generate an item insight based on content of the review; and obtaining, as output from the trained model, a review summary and the item insight.

18. The media of claim 17, wherein the operations further comprises causing the review summary to be displayed based on an identified user's interest of the item.

19. The media of claim 17, wherein the review weight further comprises a set of weights for the review based on review ratings, reviewer metadata, review dates, and review feedback.

20. The media of claim 17, wherein the trained model is a large language model.